Speaker: Xiyue Zhang (University of Bristol)

Time: 10:00 a.m. - 12:00 p.m., Mar 19, 2025, GMT+8

Venue: Rm 313, Zhihua Building, PKU

Abstract: 

As deep learning systems become increasingly integrated into high-stakes applications, ensuring their reliability and robustness is paramount. This talk explores two complementary techniques for enhancing the reliability of neural networks. First, we introduce RAC-BNN, a Risk-Averse Certification framework for Bayesian neural networks that leverages Conditional Value at Risk (CVaR) to provide probabilistic robustness guarantees under worst-case scenarios. By combining sampling-based uncertainty estimation with optimization techniques, RAC-BNN delivers tighter certified bounds and improved efficiency. Second, we present FAST, a method that enhances deep neural network testing by mitigating the over-confidence problem. FAST dynamically refines uncertainty estimation through guided feature selection, making high-confidence errors more distinguishable and improving the effectiveness of test prioritization. Together, these approaches contribute to building more resilient deep learning models, advancing both certification and testing methodologies for trustworthy AI.

Source: School of Mathematical Sciences, PKU